CTFW

我们在网络安全领域（3154 个文档）构建了一个大型带注释的程序文本数据集（CTFW）。该数据集包含有关软件漏洞分析经验的宝贵说明。我们使用我们的 LM-GNN 模型变体在多种设置中对 CTFW 进行了广泛的实验。为了展示这项任务和我们方法的普遍性，我们还尝试了来自其他两个领域（维护手册和烹饪）的程序文本，这与网络安全有很大不同。我们的实验表明，带有 BERT 句子嵌入的图卷积网络在所有三个领域都优于 BERT。We constructed a large annotated procedural text dataset named CTFW within the cybersecurity domain, which comprises 3154 documents. This dataset contains valuable explanatory content related to experiences in software vulnerability analysis. We conducted extensive experiments on CTFW across multiple experimental settings using our LM-GNN model variant. To demonstrate the generalizability of this task and our proposed method, we additionally tested procedural text samples from two other distinct domains: maintenance manuals and cooking, which are significantly different from cybersecurity. Our experimental results show that graph convolutional networks with BERT sentence embeddings outperform BERT across all three domains.

搜集汇总

背景与挑战

背景概述

CTFW是一个大型带注释的程序文本数据集，专注于网络安全领域，包含3154个文档，提供软件漏洞分析经验的宝贵说明。该数据集通过实验验证了其在不同领域的适用性，结果表明带有BERT句子嵌入的图卷积网络在网络安全、维护手册和烹饪三个领域均优于BERT。